On Fri, 8 Feb 2019 at 20:00, Richard Biener <rguent...@suse.de> wrote: > > On February 8, 2019 7:22:48 PM GMT+01:00, Jakub Jelinek <ja...@redhat.com> > wrote: > >Hi! > > > >The following testcase distilled from > >https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88739#c0 > >aborts on s390x-linux when compiled with trunk -O2 with r268332 > >reverted (or > >e.g. with -O2 and gcc 7.x) and succeeds with trunk -O2, or -O0 with any > >of > >those compilers. Tested also on x86_64-linux with -m32/-m64 make > >check. > > > >Ok for trunk? > > OK and thanks for the testcase! > > Richard. > > >2019-02-08 Jakub Jelinek <ja...@redhat.com> > > > > PR tree-optimization/88739 > > * gcc.c-torture/execute/pr88739.c: New test. > > > >--- gcc/testsuite/gcc.c-torture/execute/pr88739.c.jj 2019-01-27 > >12:44:12.526219828 +0100 > >+++ gcc/testsuite/gcc.c-torture/execute/pr88739.c 2019-02-08 > >18:48:25.880556579 +0100 > >@@ -0,0 +1,59 @@ > >+/* PR tree-optimization/88739 */ > >+#if __SIZEOF_SHORT__ == 2 && __SIZEOF_INT__ == 4 && __CHAR_BIT__ == 8 > >+struct A > >+{ > >+ unsigned int a, b, c; > >+ unsigned int d : 30; > >+ unsigned int e : 2; > >+}; > >+ > >+union U > >+{ > >+ struct A f; > >+ unsigned int g[4]; > >+ unsigned short h[8]; > >+ unsigned char i[16]; > >+}; > >+volatile union U v = { .f.d = 0x4089 }; > >+ > >+__attribute__((noipa)) void > >+bar (int x) > >+{ > >+ static int i; > >+ switch (i++) > >+ { > >+ case 0: if (x != v.f.d) __builtin_abort (); break; > >+ case 1: if (x != v.f.e) __builtin_abort (); break; > >+ case 2: if (x != v.g[3]) __builtin_abort (); break; > >+ case 3: if (x != v.h[6]) __builtin_abort (); break; > >+ case 4: if (x != v.h[7]) __builtin_abort (); break; > >+ default: __builtin_abort (); break; > >+ } > >+} > >+ > >+void > >+foo (unsigned int x) > >+{ > >+ union U u; > >+ u.f.d = x >> 2; > >+ u.f.e = 0; > >+ bar (u.f.d); > >+ bar (u.f.e); > >+ bar (u.g[3]); > >+ bar (u.h[6]); > >+ bar (u.h[7]); > >+} > >+ > >+int > >+main () > >+{ > >+ foo (0x10224); > >+ return 0; > >+} > >+#else > >+int > >+main () > >+{ > >+ return 0; > >+} > >+#endif > > > > Jakub >
Looks more elaborate than what I was deriving from #c0. I'll let you know if I see some problems on aarch64_be or armeb. Thanks, Christophe
